$$ % Typography and symbols \newcommand{\msf}[1]{\mathsf{#1}} \newcommand{\ctx}{\Gamma} \newcommand{\qamp}{&\quad} \newcommand{\qqamp}{&&\quad} \newcommand{\Coloneqq}{::=} \newcommand{\proves}{\vdash} \newcommand{\star}[1]{#1^{*}} \newcommand{\eps}{\varepsilon} \newcommand{\nul}{\varnothing} \newcommand{\brc}[1]{\{{#1}\}} \newcommand{\binopm}[2]{#1~\bar{\oplus}~#2} \newcommand{\mag}[1]{|{#1}|} \newcommand{\aequiv}{\equiv_\alpha} \newcommand{\semi}[2]{{#1};~{#2}} % Untyped lambda calculus \newcommand{\fun}[2]{\lambda ~ {#1} ~ . ~ {#2}} \newcommand{\app}[2]{#1 ~ #2} \newcommand{\fix}[3]{\msf{fix}~({#1} : {#2}) ~ . ~ #3 } \newcommand{\truet}{\msf{true}} \newcommand{\falset}{\msf{false}} \newcommand{\define}[2]{{#1} \triangleq {#2}} % Typed lambda calculus - expressions \newcommand{\funt}[3]{\lambda ~ \left(#1 : #2\right) ~ . ~ #3} \newcommand{\lett}[4]{\msf{let} ~ \hasType{#1}{#2} = #3 ~ \msf{in} ~ #4} \newcommand{\letrec}[4]{\msf{letrec} ~ \hasType{#1}{#2} = #3 ~ \msf{in} ~ #4}a \newcommand{\ift}[3]{\msf{if} ~ {#1} ~ \msf{then} ~ {#2} ~ \msf{else} ~ {#3}} \newcommand{\rec}[5]{\msf{rec}(#1; ~ #2.#3.#4)(#5)} \newcommand{\case}[5]{\msf{case} ~ {#1} ~ \{ L(#2) \to #3 \mid R(#4) \to #5 \}} \newcommand{\pair}[2]{\left({#1},~{#2}\right)} \newcommand{\proj}[2]{#1 . #2} \newcommand{\inj}[3]{\msf{inj} ~ #1 = #2 ~ \msf{as} ~ #3} \newcommand{\letv}[3]{\msf{let} ~ {#1} = {#2} ~ \msf{in} ~ {#3}} \newcommand{\fold}[2]{\msf{fold}~{#1}~\msf{as}~{#2}} \newcommand{\unfold}[1]{\msf{unfold}~{#1}} \newcommand{\poly}[2]{\Lambda~{#1}~.~ #2} \newcommand{\polyapp}[2]{{#1}~\left[{#2}\right]} \newcommand{\export}[3]{\msf{export}~ #1 ~\msf{without}~{#2}~\msf{as}~ #3} \newcommand{\import}[4]{\msf{import} ~ ({#1}, {#2}) = {#3} ~ \msf{in} ~ #4} % Typed lambda calculus - types \newcommand{\tnum}{\msf{num}} \newcommand{\tstr}{\msf{string}} \newcommand{\tint}{\msf{int}} \newcommand{\tbool}{\msf{bool}} \newcommand{\tfun}[2]{#1 \rightarrow #2} \newcommand{\tprod}[2]{#1 \times #2} \newcommand{\tsum}[2]{#1 + #2} \newcommand{\trec}[2]{\mu~{#1}~.~{#2}} \newcommand{\tvoid}{\msf{void}} \newcommand{\tunit}{\msf{unit}} \newcommand{\tpoly}[2]{\forall~{#1}~.~{#2}} \newcommand{\tmod}[2]{\exists ~ {#1} ~ . ~ #2} % WebAssembly \newcommand{\wconst}[1]{\msf{i32.const}~{#1}} \newcommand{\wbinop}[1]{\msf{i32}.{#1}} \newcommand{\wgetlocal}[1]{\msf{get\_local}~{#1}} \newcommand{\wsetlocal}[1]{\msf{set\_local}~{#1}} \newcommand{\wgetglobal}[1]{\msf{get\_global}~{#1}} \newcommand{\wsetglobal}[1]{\msf{set\_global}~{#1}} \newcommand{\wload}{\msf{i32.load}} \newcommand{\wstore}{\msf{i32.store}} \newcommand{\wsize}{\msf{memory.size}} \newcommand{\wgrow}{\msf{memory.grow}} \newcommand{\wunreachable}{\msf{unreachable}} \newcommand{\wblock}[1]{\msf{block}~{#1}} \newcommand{\wloop}[1]{\msf{loop}~{#1}} \newcommand{\wbr}[1]{\msf{br}~{#1}} \newcommand{\wbrif}[1]{\msf{br\_if}~{#1}} \newcommand{\wreturn}{\msf{return}} \newcommand{\wcall}[1]{\msf{call}~{#1}} \newcommand{\wlabel}[2]{\msf{label}~\{#1\}~{#2}} \newcommand{\wframe}[2]{\msf{frame}~({#1}, {#2})} \newcommand{\wtrapping}{\msf{trapping}} \newcommand{\wbreaking}[1]{\msf{breaking}~{#1}} \newcommand{\wreturning}[1]{\msf{returning}~{#1}} \newcommand{\wconfig}[5]{\{\msf{module}{:}~{#1};~\msf{mem}{:}~{#2};~\msf{locals}{:}~{#3};~\msf{stack}{:}~{#4};~\msf{instrs}{:}~{#5}\}} \newcommand{\wfunc}[4]{\{\msf{params}{:}~{#1};~\msf{locals}{:}~{#2};~\msf{return}~{#3};~\msf{body}{:}~{#4}\}} \newcommand{\wmodule}[1]{\{\msf{funcs}{:}~{#1}\}} \newcommand{\wcg}{\msf{globals}} \newcommand{\wcf}{\msf{funcs}} \newcommand{\wci}{\msf{instrs}} \newcommand{\wcs}{\msf{stack}} \newcommand{\wcl}{\msf{locals}} \newcommand{\wclab}{\msf{labels}} \newcommand{\wcm}{\msf{mem}} \newcommand{\wcmod}{\msf{module}} \newcommand{\wsteps}[2]{\steps{\brc{#1}}{\brc{#2}}} \newcommand{\with}{\underline{\msf{with}}} \newcommand{\wvalid}[2]{{#1} \vdash {#2}~\msf{valid}} \newcommand{\wif}[2]{\msf{if}~{#1}~{\msf{else}}~{#2}} \newcommand{\wfor}[4]{\msf{for}~(\msf{init}~{#1})~(\msf{cond}~{#2})~(\msf{post}~{#3})~{#4}} % assign4.3 custom \newcommand{\wtry}[2]{\msf{try}~{#1}~\msf{catch}~{#2}} \newcommand{\wraise}{\msf{raise}} \newcommand{\wraising}[1]{\msf{raising}~{#1}} \newcommand{\wconst}[1]{\msf{i32.const}~{#1}} \newcommand{\wbinop}[1]{\msf{i32}.{#1}} \newcommand{\wgetlocal}[1]{\msf{get\_local}~{#1}} \newcommand{\wsetlocal}[1]{\msf{set\_local}~{#1}} \newcommand{\wgetglobal}[1]{\msf{get\_global}~{#1}} \newcommand{\wsetglobal}[1]{\msf{set\_global}~{#1}} \newcommand{\wload}{\msf{i32.load}} \newcommand{\wstore}{\msf{i32.store}} \newcommand{\wsize}{\msf{memory.size}} \newcommand{\wgrow}{\msf{memory.grow}} \newcommand{\wunreachable}{\msf{unreachable}} \newcommand{\wblock}[1]{\msf{block}~{#1}} \newcommand{\wloop}[1]{\msf{loop}~{#1}} \newcommand{\wbr}[1]{\msf{br}~{#1}} \newcommand{\wbrif}[1]{\msf{br\_if}~{#1}} \newcommand{\wreturn}{\msf{return}} \newcommand{\wcall}[1]{\msf{call}~{#1}} \newcommand{\wlabel}[2]{\msf{label}~\{#1\}~{#2}} \newcommand{\wframe}[2]{\msf{frame}~({#1}, {#2})} \newcommand{\wtrapping}{\msf{trapping}} \newcommand{\wbreaking}[1]{\msf{breaking}~{#1}} \newcommand{\wreturning}[1]{\msf{returning}~{#1}} \newcommand{\wconfig}[5]{\{\msf{module}{:}~{#1};~\msf{mem}{:}~{#2};~\msf{locals}{:}~{#3};~\msf{stack}{:}~{#4};~\msf{instrs}{:}~{#5}\}} \newcommand{\wfunc}[4]{\{\msf{params}{:}~{#1};~\msf{locals}{:}~{#2};~\msf{return}~{#3};~\msf{body}{:}~{#4}\}} \newcommand{\wmodule}[1]{\{\msf{funcs}{:}~{#1}\}} \newcommand{\wcg}{\msf{globals}} \newcommand{\wcf}{\msf{funcs}} \newcommand{\wci}{\msf{instrs}} \newcommand{\wcs}{\msf{stack}} \newcommand{\wcl}{\msf{locals}} \newcommand{\wcm}{\msf{mem}} \newcommand{\wcmod}{\msf{module}} \newcommand{\wsteps}[2]{\steps{\brc{#1}}{\brc{#2}}} \newcommand{\with}{\underline{\msf{with}}} \newcommand{\wvalid}[2]{{#1} \vdash {#2}~\msf{valid}} % assign4.3 custom \newcommand{\wtry}[2]{\msf{try}~{#1}~\msf{catch}~{#2}} \newcommand{\wraise}{\msf{raise}} \newcommand{\wraising}[1]{\msf{raising}~{#1}} \newcommand{\wif}[2]{\msf{if}~{#1}~{\msf{else}}~{#2}} \newcommand{\wfor}[4]{\msf{for}~(\msf{init}~{#1})~(\msf{cond}~{#2})~(\msf{post}~{#3})~{#4}} \newcommand{\windirect}[1]{\msf{call\_indirect}~{#1}} % session types \newcommand{\ssend}[2]{\msf{send}~{#1};~{#2}} \newcommand{\srecv}[2]{\msf{recv}~{#1};~{#2}} \newcommand{\soffer}[4]{\msf{offer}~\{{#1}\colon({#2})\mid{#3}\colon({#4})\}} \newcommand{\schoose}[4]{\msf{choose}~\{{#1}\colon({#2})\mid{#3}\colon({#4})\}} \newcommand{\srec}[1]{\msf{label};~{#1}} \newcommand{\sgoto}[1]{\msf{goto}~{#1}} \newcommand{\dual}[1]{\overline{#1}} % Inference rules \newcommand{\inferrule}[3][]{\cfrac{#2}{#3}\;{#1}} \newcommand{\ir}[3]{\inferrule[\text{(#1)}]{#2}{#3}} \newcommand{\s}{\hspace{1em}} \newcommand{\nl}{\\[2em]} \newcommand{\evalto}{\boldsymbol{\overset{*}{\mapsto}}} \newcommand{\steps}[2]{#1 \boldsymbol{\mapsto} #2} \newcommand{\evals}[2]{#1 \evalto #2} \newcommand{\subst}[3]{[#1 \rightarrow #2] ~ #3} \newcommand{\dynJ}[2]{#1 \proves #2} \newcommand{\dynJC}[1]{\dynJ{\ctx}{#1}} \newcommand{\typeJ}[3]{#1 \proves \hasType{#2}{#3}} \newcommand{\typeJC}[2]{\typeJ{\ctx}{#1}{#2}} \newcommand{\hasType}[2]{#1 : #2} \newcommand{\val}[1]{#1~\msf{val}} \newcommand{\num}[1]{\msf{Int}(#1)} \newcommand{\err}[1]{#1~\msf{err}} \newcommand{\trans}[2]{#1 \leadsto #2} \newcommand{\size}[1]{\left|#1\right|} $$

&Notepad

Types Over Strings: Extensible Architectures in Rust

Will Crichton   —   April 12, 2020
I show how to implement event listeners and dependency injection using Rust's type system to avoid common errors in stringly-typed versions of these extensible architectures.

All code in this note is available on Github.

Types are a useful tool to make sure software libraries work together. I expect an int, you give me a string, compiler raises an error. In my experience, types, interfaces and encapsulation work best when using black-box APIs: hash maps, regexes, HTTP requests. The programmer controls the top-level program, i.e. the types that orchestrate the individual pieces.

However, the Type Life tends to become harder when dealing with frameworks, or any kind of extensible architecture where the programmer is plugging components into a bigger system they don’t control. For example, consider the venerable EventTarget.addEventListener from JavaScript:

document.body.addEventListener('click', function(event) {
  console.log(event.clientX, event.clientY);
});

This is an extensible architecture in the sense that you don’t control the DOM API, nor can you control what events are registered to DOM nodes. That changes all the. You can just hook-in to events the browser provides.

Let’s look at this function with two related questions: what is the type of addEventListener? And what kinds of errors can we make using this function?

The most basic type for addEventListener(event, listener) is that event is a string, and listener is a function. A function from what to what? The input is an “event”, whatever that means, and there is no output. Indeed, when addEventListener gets ported into a statically typed language like ReasonML, the type definition looks like:

external addEventListener : (string, Dom.event => unit) => unit;

Here’s two ways we can mess this up.

In this note, I’ll show how you can avoid errors like these by designing type-safe extensible systems. We’ll look at two examples: event handling and dependency injection. The code will use Rust, but the lessons apply to any functional language.

Type-safe event listeners

The basic idea in making type-safe extensible architectures is to replace strings with types. Any time you use a string as an identifier (e.g. for an event, or a software component), use a type instead. Why?

Here’s an example of using a type-safe event API in Rust.

struct OnClick {
  mouse_x: i32,
  mouse_y: i32,
}

let mut dispatcher = EventDispatcher::new();

dispatcher.add_event_listener(|event: &OnClick| {
  assert_eq!(event.mouse_x, 10);
  assert_eq!(event.mouse_y, 5);
});

dispatcher.trigger(&OnClick {
  mouse_x: 10,
  mouse_y: 5,
})

In the snippet above, the event parameter E is inferred from the type parameter OnClick of the closure, rather than using a string argument. This solves both of our errors:

Stupendous! But how does it work? Here’s the basic structure.

// Events must not contain pointers to things, for simplicity
trait Event: 'static {}

// An event listener is a function from an event to nil
trait EventListener<E: Event> = FnMut(&E) -> () + 'static;

// An event dispatcher holds all the event listeners
struct EventDispatcher { /* .. */ }
impl EventDispatcher {
  // Registers a function `f` to listen for an event `E`
  fn add_event_listener<E, F>(&mut self, f: F)
  where
    E: Event,
    F: EventListener<E>
  {
    /* .. */
  }

  // Runs all the registered listeners for the event `E`
  fn trigger<E>(&mut self, event: &E)
  where
    E: Event
  {
    /* .. */
  }
}

In order to implement EventDispatcher, we need a data structure that can hold all the listeners. Given an event E, it should provide all the listeners for E. This raises two questions:

Mapping types to values

We will briefly detour to make a critical building block: the TypeMap. Based on our two requirements above, we will make a data structure that a) maps types to values, and b) holds values of different types.

Rust has std::any for this purpose. TypeId allows us to get a unique, hashable identifier for each type. Any allows us to up-cast/down-cast objects at runtime. Hence, our TypeMap will map from TypeId to Box<dyn Any>.

use std::collections::HashMap;
use std::any::{TypeId, Any};

struct TypeMap(HashMap<TypeId, Box<dyn Any>>);

To add an element to the map:

impl TypeMap {
  pub fn set<T: Any + 'static>(&mut self, t: T) {
    self.0.insert(TypeId::of::<T>(), Box::new(t));
  }
}

Aside: the syntax ::<OnClick> is Rust’s “turbofish”. It explicitly binds a type parameter of a polymorphic function, rather than leaving it to be inferred. Further explanation here.

This means our map has one unique value for a given type. For example, if we use the TypeMap like this:

let mut map = TypeMap::new();
map.set::<i32>(1);

Then we insert a value 1 at the key TypeId::of::<i32>(). We can also implement has and get functions:

impl TypeMap {
  pub fn has<T: 'static+Any>(&self) -> bool {
    self.0.contains_key(&TypeId::of::<T>())
  }

  pub fn get_mut<T: 'static+Any>(&mut self) -> Option<&mut T> {
    self.0.get_mut(&TypeId::of::<T>()).map(|t| {
      t.downcast_mut::<T>().unwrap()
    })
  }
}

Look carefully at get_mut. The inner hash map returns a value of type Box<dyn Any>, which we can downcast_mut to become a value of type &mut T. This operation is guaranteed to not fail, because only values of type T are stored in the hash map under the key for T.

Finishing our event system

With the TypeMap in hand, we can finish our event system. For the EventDispatcher, the TypeMap will map from events to a vector of listeners.

pub struct EventDispatcher(TypeMap);

// Type alias for a list of listeners for an event
type ListenerVec<E> = Vec<Box<dyn EventListener<E>>>;

impl EventDispatcher {
  pub fn add_event_listener<E, F>(&mut self, f: F)
  where
    E: Event,
    F: EventListener<E>,
  {
    if !self.0.has::<ListenerVec<E>>() {
      self.0.set::<ListenerVec<E>>(Vec::new());
    }

    let listeners = self.0.get_mut::<ListenerVec<E>>().unwrap();
    listeners.push(Box::new(f));
  }
}

When asked to register an event listener, we get the ListenerVec from the TypeMap and insert the new listener. Trigger is similarly simple:

impl EventDispatcher {
  pub fn trigger<E: Event>(&mut self, event: &E) {
    if let Some(listeners) = self.0.get_mut::<ListenerVec<E>>() {
      for callback in listeners {
        callback(event);
      }
    }
  }
}

That’s it! Now we have a type-safe event system.

Type-safe dependency injection

To avoid the impression that this pattern is specific to event listeners, I want to show another example of using types over strings in extensible architectures. I started playing with specs recently, an entity-component-system architecture written in Rust. Its System uses a dependency-injection-like pattern, so I wanted to distill that to a pedagogical example here.

Dependency injection is often found rife with XML file, string keys, and other issues akin to those with event systems. Hence, it makes for a good use case here.

Motivation

The basic idea of dependency injection (DI) is that you have a component that depends on another, like a web server using a database. However, you don’t want to hard-code a particular database constructor, and rather make it easy to swap in-and-out different databases. For example:

trait Database {
  fn name(&self) -> &'static str;
}

struct MySQL;
impl Database for MySQL {
  fn name(&self) -> &'static str { "MySQL" }
}

struct Postgres;
impl Database for Postgres {
  fn name(&self) -> &'static str { "Postgres" }
}

struct WebServer { db: Box<dyn Database> }
impl WebServer {
  fn run(&self) {
    println!("Db name: {}", self.db.name());
  }
}

To implement DI, we need two things:

With these pieces, we can use our DI system like so:

let mut manager = DIManager::new();
manager.build::<MySQL>().unwrap();
let server = manager.build::<WebServer>().unwrap();
server.lock().unwrap().run(); // prints Db name: MySQL

DI constructors

First, we’ll define a trait DIBuilder that represents a constructor within our DI system.

trait DIBuilder {
  type Input;
  type Output;

  fn build(input: Self::Input) -> Self::Output;
}

The build method is a static method (doesn’t take self as input). It just takes Input as input, and produces Output as output. The key idea is that because Input and Output are associated types, we can inspect them later on. We will need to find values for Input and to store Output in our DI manager.

We implement DIBuilder for each type in the system. The databases have no inputs, so their input is (). Their return type is Box<dyn Database>, meaning they are explicitly cast to a trait object so they can be used interchangeably.

impl DIBuilder for MySQL {
  type Input = ();
  type Output = Box<dyn Database>;
  fn build((): ()) -> Box<dyn Database> {
    Box::new(MySQL)
  }
}

impl DIBuilder for Postgres {
  type Input = ();
  type Output = Box<dyn Database>;
  fn build((): ()) -> Box<dyn Database> {
    Box::new(Postgres)
  }
}

impl DIBuilder for WebServer {
  type Input = (Box<dyn Database>,);
  type Output = WebServer;

  fn build((db,): Self::Input) -> WebServer {
    WebServer { db }
  }
}

DI manager

Now that we know the dependency structure of our objects, we need a centralized manager to store the objects and fetch their dependencies.

struct DIManager(TypeMap);

In this TypeMap, we will store the constructed objects. For example, once we make a Box<dyn Database>, then we will map TypeId::of::<Box<dyn Database>> to one of Box<Postgres> or Box<MySQL>.

impl DIManager {
  fn build<T: DIBuilder>(&mut self) -> Option<T::Output> {
    let input = /* get the inputs, somehow */;
    let obj = T::build(input);
    self.0.set::<T::Output>(obj);
    Some(obj)
  }
}

Ignoring how we fetch dependencies for now, this function calls the DIBuilder::build implementation for T, then stores the result in the TypeMap. This approach almost works, except not for Rust: ownership of obj is passed into TypeMap and the result of build.

And, intuitively, this makes sense. If a component like a database cursor needs to be shared across many downstream components, it needs some kind of access protection. Hence, we tweak our interface a bit to wrap everything in an Arc and Mutex.

type DIObj<T> = Arc<Mutex<T>>;

impl DIManager {
  fn build<T: DIBuilder>(&mut self) -> Option<DIObj<T::Output>> {
    let input = /* get the inputs, somehow */;
    let obj = T::build(input);
    let sync_obj = Arc::new(Mutex::new(obj));
    self.0.set::<DIObj<T::Output>>(sync_obj.clone());
    Some(sync_obj)
  }
}

DI dependencies

Finally, we need a way to implement the let input line in DIManager::build. Given a type T: DIBuilder, it has an associated type T::Input that represents the inputs needed to build it.

To simplify the problem, imagine T::Input = S where S: DIBuilder. Then if S has already been built, e.g. T = WebServer and S = Box<dyn Database>, we can fetch it directly from the typemap:

let input = self.0.get::<T::Input>().map(|obj| obj.clone())?;

However, in practice T::Input could be several dependencies. For example, if our server depends on a configuration, it might be:

impl DIBuilder for WebServer {
  type Input = (Box<dyn Database>, Box<dyn ServerConfig>);
  type Output = WebServer;

  fn build((db, config): Self::Input) -> WebServer {
    WebServer { db, config }
  }
}

Let’s assume now T::Input is always a tuple (S1, S2, ...) of types where each type Si : DIBuilder. Ideally, we could write something like:

let input = (T::Input).map(|S| {
  self.0.get::<S>().map(|obj| obj.clone()).unwrap()
});

But, alas, our language of expressions is not our language of types. Such a thing is the provenance of languages we can only dream about. Instead, we have to cleverly use traits to inductively define a way to extract inputs from the tuple. To start, we’ll make a trait that gets an object of a particular type from the DIManager:

trait GetInput: Sized {
  fn get_input(manager: &DIManager) -> Option<Self>;
}

For DIObj<T>, this means looking up the type in the TypeMap:

impl<T: 'static> GetInput for DIObj<T> {
  fn get_input(manager: &DIManager) -> Option<Self> {
    manager.0.get::<Self>().map(|obj| obj.clone())
  }
}

Then for tuples of DIObj<T>, we can make an inductive definition like so:

impl GetInput for () {
  fn get_input(_manager: &DIManager) -> Option<Self> {
    Some(())
  }
}

impl<S: GetInput, T: GetInput> GetInput for (S, T) {
  fn get_input(manager: &DIManager) -> Option<Self> {
    S::get_input(manager).and_then(|s| {
      T::get_input(manager).and_then(|t| {
        Some((s, t))
      })
    })
  }
}

Then we can modify our WebServer example to use an inductive structure:

impl DIBuilder for WebServer {
  type Input = (DIObj<Box<dyn Database>>,(DIObj<Box<dyn ServerConfig>>,()));
  type Output = WebServer;

  fn build((db, (config, ())): Self::Input) -> WebServer {
    WebServer { db, config }
  }
}

Aside: in practice, rather than doing nested pairs, you can use a macro to create the GetInput impl for many tuple types like this.

And at last, we can implement DIManager::build:

impl DIManager {
  pub fn build<T: DIBuilder>(&mut self) -> Option<DIObj<T::Output>> {
    let input = T::Input::get_input(self)?;
    let obj = T::build(input);
    let sync_obj = Arc::new(Mutex::new(obj));
    self.0.set::<DIObj<T::Output>>(sync_obj.clone());
    Some(sync_obj)
  }
}

Now, the expression T::Input::get_input(self) will convert a tuple of types into a tuple of values of those types.

Final API example

With the API complete, our example now looks like this:

let mut manager = DIManager::new();
manager.build::<MySQL>().unwrap();
let server = manager.build::<WebServer>().unwrap();
server.lock().unwrap().run();

We can construct a WebServer without explicitly passing in a dyn Database instance. When we use the server, we have to explicitly call .lock() now that it’s wrapped in a mutex. And lo, our dependencies have been injected.

Conclusion

Across these examples, there are a few main language-level mechanisms that enable the high-level patterns:

It seems that these mechanisms are important for type-safe extensible architectures. I’m super excited to see the next generation of frameworks like specs and Rocket creatively applying these ideas to make programming safer.